Files
foxhunt/config
jgrusewski d6bd7e1e00 fix: all 11 smoke tests pass — smoketest generalization tuning
Smoketest TOML: added [generalization] section with toned-down params
for tiny [64,64] network:
- anti_lr mults 1.1x/0.9x (vs production 3x/0.3x)
- stochastic_depth 10% (vs 20%)
- feature_mask 15% (vs 30%)
- pruning_epoch 100 (beyond 50-epoch test)
- saboteur warmup 100 (beyond test)
- bottleneck disabled (tiny network)

50-epoch tests: set gradient_collapse_patience=50000 to prevent
false positive from deferred grad_norm readback (per-step counter
sees 0 because grad_norm_finalize runs outside graph).

CUDA context leak fix: experience collector loads EXPERIENCE_KERNELS_CUBIN
once for all domain_rand + saboteur kernels (was 4 separate loads).

Configurable pruning: pruning_epoch and pruning_fraction moved from
hardcoded to DQNHyperparameters + GpuDqnTrainConfig + TOML.

Test results:
- ml unit: 893 passed, 2 failed (DB only)
- ml-dqn unit: 359 passed, 0 failed
- Smoke tests: 11/11 PASS
  - 3-epoch: Sharpe 9.34
  - 50-epoch convergence: Sharpe 0.95, MaxDD 11.1%, +63% return
  - 50-epoch walk-forward: best Sharpe 8.13

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-30 14:10:48 +02:00
..